Complexity adaptation in video encoders for power limited platforms

With the emergence of video services on power limited platforms, it is necessary to consider both performance-centric and constraint-centric signal processing techniques. Traditionally, video applications have a bandwidth or computational resources constraint or both. The recent H.264/AVC video compression standard offers significantly improved efficiency and flexibility compared to previous standards, which leads to less emphasis on bandwidth. However, its high computational complexity is a problem for codecs running on power limited plat- forms. Therefore, a technique that integrates both complexity and bandwidth issues in a single framework should be considered. In this thesis we investigate complexity adaptation of a video coder which focuses on managing computational complexity and provides significant complexity savings when applied to recent standards. It consists of three sub functions specially designed for reducing complexity and a framework for using these sub functions; Variable Block Size (VBS) partitioning, fast motion estimation, skip macroblock detection, and complexity adaptation framework. Firstly, the VBS partitioning algorithm based on the Walsh Hadamard Transform (WHT) is presented. The key idea is to segment regions of an image as edges or flat regions based on the fact that prediction errors are mainly affected by edges. Secondly, a fast motion estimation algorithm called Fast Walsh Boundary Search (FWBS) is presented on the VBS partitioned images. Its results outperform other commonly used fast algorithms. Thirdly, a skip macroblock detection algorithm is proposed for use prior to motion estimation by estimating the Discrete Cosine Transform (DCT) coefficients after quantisation. A new orthogonal transform called the S-transform is presented for predicting Integer DCT coefficients from Walsh Hadamard Transform coefficients. Complexity saving is achieved by deciding which macroblocks need to be processed and which can be skipped without processing. Simulation results show that the proposed algorithm achieves significant complexity savings with a negligible loss in rate-distortion performance. Finally, a complexity adaptation framework which combines all three techniques mentioned above is proposed for maximizing the perceptual quality of coded video on a complexity constrained platform